Using AI to fight eavesdropping

The situation of “eavesdropping” happens a lot in today’s era. It is not difficult for internet users to come across a series of ads on Facebook, YouTube… about the type of product they mentioned not long ago in conversation with others.

That is the “accomplishment” of spyware. Turning on loud music or using the sound of running water in the background is one of the common ways to prevent eavesdropping, but it is often distracting.

Recently, researchers at Columbia University (USA) have created a breakthrough technology called “Neural Voice Camouflage”, which can emit noise in any context, helping to prevent devices from entering the environment. intelligently track users. This tool generates a custom audio cue for the background of the conversation, which can confuse eavesdropping artificial intelligence (AI) devices.

Many smart devices have built-in eavesdropping software Photo: ALAMY

The “Neural Voice Camouflage” technology uses machine learning (a subfield of computer science that is capable of self-learning based on input without needing to be specifically programmed), in where algorithms figure out data patterns to tune the sound in such a way that the AI ​​confuses the user’s voice for something else.

Simply put, using one AI to fool another. However, this process is not easy. The machine learning AI needs to process the entire audio track before jamming, which won’t work if the user wants to use the tool to avoid real-time eavesdropping.

That’s why the researchers set up a neural network – a brain-inspired machine learning system – to effectively predict the future. They trained it through hours of voice recording so that it could continuously process 2-second audio clips and mask the words that might come next.

According to Science Magazine (USA), the system also helps to increase the word error rate of automatic speech recognition software (ASR) from 11.3% to 80.2%. The work was presented last month at the ICLR 2022 Conference on Machine Learning and Artificial Intelligence in an online format.

Carl Vondrick, a computer science expert at Columbia University, says it’s important that all functions work fast enough. Mr. Vondrick said: “Our algorithm is capable of blocking an eavesdropping microphone that accurately collects the user’s speech about 80% of the time. In testing, it was the fastest and most accurate tool. “.

The expert explains that the aforementioned anti-spy tool works even if the user is not aware of anything about the eavesdropping device, such as their location or even the computer software with the eavesdropping feature. slip.

Basically, the new tool helps to disguise a person’s voice through the air, protect the conversation from eavesdropping systems and not inconvenience conversations between people in the room.